agent-relay-server 0.121.2 → 0.121.4

package/docs/openapi.json

@@ -2,7 +2,7 @@
   "openapi": "3.1.0",
   "info": {
     "title": "Agent Relay API",
-    "version": "0.121.2",
+    "version": "0.121.4",
     "description": "Real-time message bus for inter-agent communication. Agent-first: this spec is designed for machine consumption — agents can self-discover the full API surface via GET /api/spec.",
     "license": {
       "name": "MIT",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-server",
-  "version": "0.121.2",
+  "version": "0.121.4",
   "description": "Lightweight HTTP message relay for inter-agent communication across machines",
   "module": "src/index.ts",
   "type": "module",
@@ -37,7 +37,7 @@
     "CONTRIBUTING.md"
   ],
   "dependencies": {
-    "agent-relay-channels-host": "0.121.2",
+    "agent-relay-channels-host": "0.121.4",
     "agent-relay-providers": "0.104.3",
     "agent-relay-sdk": "0.2.113",
     "ajv": "^8.20.0"

package/runner/plugins/claude/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "agent-relay-runner",
   "description": "Thin Agent Relay runner bridge for Claude Code",
-  "version": "0.121.2",
+  "version": "0.121.4",
   "agentRelayContracts": {
     "providerPluginProtocol": 1
   }

package/src/bus.ts

@@ -174,7 +174,12 @@ function handleFrame(ws: BusWebSocket, frame: ReturnType<typeof validateClientFr
               // frame. Flush the held report-up on THIS edge too, else `before.status` is already
               // non-busy at the next status frame → the flush never fires → the child deadlocks idle
               // with its verdict held and the coordinator waits forever.
-              if (current?.status === "busy" && frame.payload.status !== "busy") flushIdleLineageReport(conn.agentId);
+              // #1037 (re-regression) — gate on a genuine at-rest `idle`, NOT merely "not busy". A
+              // real turn-end always reports `idle`; a transient non-idle status (a stray `online`,
+              // a malformed frame) is NOT the end of the child's work and must not flush its held
+              // mid-work capture upstream. Terminal exit edges (WS-close/stale sweep) keep their
+              // unconditional flush for the #1040 delivery guarantee — this only tightens the LIVE edge.
+              if (current?.status === "busy" && frame.payload.status === "idle") flushIdleLineageReport(conn.agentId);
             }
             emitAgentStatusEvent(conn.agentId);
             // #416 — flush queued messages if the agent transitioned from unavailable to available.
@@ -220,8 +225,11 @@ function handleFrame(ws: BusWebSocket, frame: ReturnType<typeof validateClientFr
         // #237 stop-hook → instant ⚪↔🟡 changes detection on the turn-end edge.
         probeWorkspaceOnTurnEnd(before, after);
         // #1037 — the same busy→idle turn-end edge flushes a quiet child's deferred turn-final
-        // report-up to its spawner (held back while the child was mid-work).
-        if (before?.status === "busy" && after && after.status !== "busy") flushIdleLineageReport(conn.agentId);
+        // report-up to its spawner (held back while the child was mid-work). Gate on a genuine
+        // at-rest `idle`: a real turn-end reports `idle`, whereas a transient non-idle status
+        // (stray `online`, reconnect) is not the child going to rest and must not flush its held
+        // mid-work capture. Terminal exit edges keep the #1040 unconditional flush.
+        if (before?.status === "busy" && after && after.status === "idle") flushIdleLineageReport(conn.agentId);
         // #633 — silent clean-exit: a runner that stays alive after its provider terminally exited
         // marks the offline edge with `terminalProviderExit`; the owning service turns that into a
         // terminal event + #636 diagnosis (and clears the hold on a genuine resume).

package/src/db/messages.ts

@@ -347,6 +347,56 @@ function childIsBusy(child: AgentCard): boolean {
   return child.status === "busy";
 }
 
+// #1037 (re-regression) — the report-up must reach the spawner as a CONCISE result, not the full
+// turn transcript. In `chatCaptureMode: "full"` (and for rambling multi-block recaps) the captured
+// turn-final body is the ENTIRE turn's narration — every "Let me read X… now Y… now Z…" segment
+// joined by the canonical `\n\n` block boundary (runner extractLastAssistantTurn). Forwarding it
+// verbatim floods the coordinator with a wall of mid-work text (the exact bug #1037 keeps
+// re-shipping). Idle-gating alone can't help: this is a genuine turn-final at a genuine idle, so
+// the payload itself must be distilled. Short bodies pass through untouched (concise finals are
+// unaffected). Pure function of the body → deterministic, so re-promotion (the idle flush) yields
+// the identical distilled body + idempotency key and dedups cleanly.
+//
+// #1056 — cut on a SEMANTIC result boundary, not a block count. The old heuristic kept the last N
+// `\n\n`-separated blocks by a size budget; that is fragile to how a model CHUNKS its answer. A
+// chatty Claude emitted its `## SUMMARY` result as one block (kept whole), but a chatty Codex
+// paragraph-broke the heading away from its bullets, so the tail-block budget clipped the header +
+// the first bullet — distillation ate the TOP of the result, delivering it mid-sentence. Fix: keep
+// everything from the FIRST markdown heading onward. Workers conclude under a heading (`## SUMMARY`,
+// `## FINAL REPORT`), narration is prose chatter emitted before it — so the result is the trailing
+// region a heading marks the start of, and cutting at the heading drops the leading narration while
+// keeping the WHOLE result, however the model split it. We cut at the FIRST heading (not the last):
+// the result may itself contain sub-headings (`### Details`) or multiple `##` sections, and "last
+// heading" would slice into it — the very clip #1056 is about; "first heading" never can, since it
+// sits at or before the result's start. Fallback (no heading anywhere — rare; workers are told to
+// emit one): there is no reliable boundary, so keep the trailing answer and trim only the leading
+// narration flood by the legacy tail budget, never slicing the block we keep.
+const REPORT_UP_DISTILL_THRESHOLD = 800;
+const REPORT_UP_TAIL_MAX = 800;
+const REPORT_UP_ELISION = "[…mid-work narration trimmed by report-up…]";
+const REPORT_UP_HEADING = /^#{1,6}\s/m;
+
+function distillLineageReportBody(body: string): string {
+  const trimmed = body.trim();
+  if (trimmed.length <= REPORT_UP_DISTILL_THRESHOLD) return body;
+  const heading = REPORT_UP_HEADING.exec(trimmed);
+  if (heading) return heading.index === 0 ? body : `${REPORT_UP_ELISION}\n\n${trimmed.slice(heading.index).trim()}`;
+  const blocks = trimmed.split(/\n{2,}/).map((block) => block.trim()).filter(Boolean);
+  const kept: string[] = [];
+  let size = 0;
+  for (let i = blocks.length - 1; i >= 0; i--) {
+    const block = blocks[i]!;
+    if (kept.length > 0 && size + block.length > REPORT_UP_TAIL_MAX) break;
+    kept.unshift(block);
+    size += block.length + 2;
+  }
+  let tail = kept.join("\n\n");
+  // A single oversized trailing block (unsegmented narration) — hard-cap to its own tail.
+  if (tail.length > REPORT_UP_TAIL_MAX) tail = tail.slice(tail.length - REPORT_UP_TAIL_MAX).replace(/^\S*\s/, "").trimStart();
+  if (tail.length >= trimmed.length) return body; // nothing meaningful trimmed
+  return `${REPORT_UP_ELISION}\n\n${tail}`;
+}
+
 // #1037 — a quiet child's turn-final response that arrives while the child is still busy is a
 // mid-task yield, not the end of its work. Instead of shipping it upstream (mid-work narration) or
 // dropping it (a genuine last turn whose idle status simply hasn't landed yet), we HOLD the latest
@@ -493,7 +543,9 @@ export function promoteLineageCapturedResponse(message: Message): { message: Mes
     from: message.from,
     to: parent,
     kind: "chat",
-    body: message.body,
+    // #1037 — distill the turn-final report-up to a concise result; verbose child-stream captures
+    // (an explicit subscriber that asked for the raw stream) are forwarded untouched.
+    body: stream === "responses" ? distillLineageReportBody(message.body) : message.body,
     replyExpected: false,
     idempotencyKey,
     payload: stream === "responses"